Lightweight Simulation Scripting with Proto

vroomhuhSoftware and s/w Development

Nov 4, 2013 (3 years and 7 months ago)

57 views

Lightweight Simulation Scripting with Proto
Jacob Beal
Raytheon BBN Technologies
Cambridge,MA,USA,02138
Email:jakebeal@bbn.com
Kyle Usbeck
Raytheon BBN Technologies
Cambridge,MA,USA,02138
Email:kusbeck@bbn.com
Brian Krisler
Raytheon BBN Technologies
Cambridge,MA,USA,02138
Email:bkrisler@bbn.com
Abstract—Modern game engines make it easy to create com-
plex realistic environments for entertainment or for training,but
scripting the behavior of agents in these environments is still a
major challenge.Spatial computing languages such as Proto [1]
provide a possible solution,but need to be adapted for practical
scripting use.We have begun to address this problem by linking
Proto with the Unity game engine and by creating a Proto library
for scripting the behavior of groups of agents.We validate our
approach by demonstrating compact scripting of three complex
agent interaction scenarios.
I.INTRODUCTION
Modern game engines [2],provide the core components
necessary to quickly produce simulations that just a few years
ago required complex,custom solutions.The proliferation of
these generic engines has led to the emergence of a new
category of games,referred to as serious games [3].The main
focus area for serious games is training,where systems such as
the US Navy VESSEL trainer [4] are used to reduce classroom
lecture times and promote active learning.
Every game engine has a scripting environment that pro-
vides a language and core API for customizing interactions
with and within the game.While these APIs are typically
robust and allow for complete control of all objects within
the game world,they are seriously limited in their support for
quickly scripting behaviors for large groups of autonomous
agents.For example,in the creation of a training game where
a trainee would have to function in a large crowd,providing
the movement flow and heterogeneous interactions typical of a
realistic crowd would require many complex pieces of custom
code,perhaps down to the level of individual agents.This
requirement limits the inclusion of many autonomous agents
in a training scenario.
In this paper,we address this problem by linking the
Proto spatial computing language to Unity [5],a widely used
modern game engine.We then create a library for scripting
the behavior of groups of agents and demonstrate how our
approach allows compact scripting for large groups of agents
in a realistic simulation environment.
II.BACKGROUND
Although there has been much previous work on agent be-
havior programming and simulation,there are significant gaps
Work partially sponsored by DARPA;the views and conclusions contained
in this document are those of the authors and not DARPA or the U.S.
Government.
in the capabilities of existing solutions.These current solutions
can be classified into three categories:single-agent behavioral
models,multi-agent toolkits,and spatial computing platforms.
Adetailed reviewdiscussing many of the approaches described
in this section can be found in [6].
Many game engines simply use conventional programming
languages (or their own domain-specific variants) for their
scripting languages.For example,Unity uses scripts written
in JavaScript,C#,and Boo.More sophisticated single-agent
behavioral models include conceptual models of agent be-
havior and agent frameworks (for implementing agent behav-
iors).Conceptual behavior models,such as the Belief-Desire-
Intent (BDI) agent model [7],offer high-level descriptions
of agent internals.Agent frameworks (often called “agent
architectures”) described thoroughly in [8] and [9],provide
tools (e.g.,agent administration,messaging,mobility,logging,
etc.) for implementing agent behaviors.These frameworks
and behavioral models,however,rarely provide aggregate
programming/modeling tools that are useful for MAS control.
Multi-agent System (MAS) modeling and simulation tool-
kits tend to focus on interactions:both inter-agent interactions
and interactions between agents and their environment.MAS
modeling and simulation toolkits include languages for mod-
eling the MAS,and tools for simulating the running agent
system.For example,NetLogo [10] extends the Logo language
to allow agent coordination and provides a graphical tool for
simulating the agent behaviors.Most existing MAS model-
ing and simulation toolkits lack realism in their simulation
environments,and therefore do not provide language features
for realistic behavior.Furthermore,most MAS toolkits lack
features for spatial aggregate programming,which we describe
next,which enable scalable descriptions of aggregate behavior
(i.e.,the number of agents don’t need to be specified a priori).
A more-recent approach is spatial computing,which as-
sumes communication is constrained to agents near one an-
other in space.The implication of this assumption is that it be-
comes necessary to consider the spatial structure of the system
in planning the solution.Proto [1] is a purely-functional LISP-
like language that is designed with spatial constructs (i.e.,
operations to measure and manipulate space-time,compute
spatial patterns,and evolve dynamically).General purpose
spatial languages such as Proto or MGS [11] are capable of
elegantly and concisely describing aggregate MAS behavior
(sometimes labeled “emergent behavior”) [12],but often have
unusual programming models.For example,Proto is a purely
functional language and does not offer the imperative-style
MAS scripting that is familiar to game-based agent behavior
developers.Furthermore,the simulators used for running and
testing spatial languages tend to lack the realism that is
available from recent physics simulators and game engines.
III.APPROACH
Our approach for creating a scalable aggregate scripting
language for realistic simulation environment has three com-
ponents:(1) the Unity game engine [5] provides realism in
the simulation environment (e.g.,terrain modeling,realistic
physics simulation,entity modeling),(2) the Proto spatial
computing language provides constructs for scalable aggregate
programming,and (3) agent behavior scripting is facilitated by
a novel Proto library comprising group behavior primitives and
a novel macro library for imperative-style scripting.
A.Connecting Proto and Unity
Proto has three main components:(1) the spatial language,
(2) a global-to-local compiler which accepts a global behavior
description (in Proto language) and outputs a virtual machine
(VM) binary for the Proto VM,and (3) the VMthat interprets
Proto VM instructions on each device.In order to be able to
execute global Proto programs from within the realistic Unity
simulation environment,we first make the Proto compiler in-
vokable from within Unity.Next,we create an implementation
of the Proto VMfor reading information from and performing
operations upon Unity agents.Finally,we create an interface
for developers to control parameters of the Proto-Unity plugin.
1) Invoking the Proto Compiler:Proto uses a global-to-
local compiler to convert global behavior descriptions into
local (i.e.,per-device) programs.It is important that this
compiler be integrated into the final solution so that end-users
can write programs for groups of agents within the modeling
and simulation toolkit.
The reference implementation of the Proto compiler is
written in C++ and Unity has a C/C++ API for its plugins,so
one option would be to directly integrate the Proto compiler
into a Unity plugin.This would have required maintaining a
branch of the Proto compiler with a Unity-friendly interface,
however.Instead,we created a simpler plugin that invokes
an external installation of the standard Proto compiler.Thus,
Unity,supplied with a Proto program,invokes the external
Proto compiler on that programand receives in return the Proto
VM instructions (a.k.a.,Proto opcodes) that specify how each
agent should act.
2) A Proto VM Implementation for Unity:Next,we need
a mechanism for controlling agents within Unity according to
the behaviors described by the local Proto VM instructions.
The Proto reference implementation already contains a VM
suitable for most environments—requiring the developer to
implement only a small set of platform-specific functions (e.g.,
how the machine allocates memory,broadcasts messages to
other devices,etc.).Likewise,the continuous time model of
Proto programs means there is no problem matching simula-
tion rates:the VM execution rate can simply be derived from
its Unity environment.Importing the Proto VM to Unity was
not significantly different or more difficult than prior imports
on various embedded platforms:to do so,we constructed a
Unity plugin that implements the required platform-specific
functions using tools from the Unity API.For example,one
such function uses the Unity utility for computing the distance
between Unity agents to implement a unit-disc communication
model.Of course,this model could be extended to incorporate
other information available from Unity (e.g.,line-of-sight) for
improved realism.
3) An Interface to the Proto-Unity Plugin:Finally,we
created an interface for controlling the Proto-Unity plugin.
This engineering interface is not meant for end-users,but
instead is designed to help agent script developers by providing
functionality similar to that of the reference implementation
of the Proto simulator.For example,the interface can show
the network topology of the Unity agents by drawing lines
between the agents within communication range and allows the
developer to change the devices’ communication radii on-the-
fly.This allows a developer to “tweak” simulation parameters
during development,then set them and remove the interface
when the simulation is finished and provided to users.
B.Group Behavior Primitives
We next need a library of “primitives” for group behaviors—
simple ways of describing what we want a collection of agents
to do.These will be the basis for the agent scripts that we
build.We build this library after the fashion of [13],as Proto
functions that compute vector fields for the desired motion of
agents.We can then produce complex behaviors by mixing
these vector fields together in various ways.
We have created an initial library of eight behaviors.Fig-
ure 1 shows examples of these behaviors being applied to
agents in Unity using the Proto/Unity bridge.Here we present
only the API for the behaviors;their implementation is similar
(or in some cases identical) to code presented in [13].
(random-walk)
Parameter
Type
Description
RETURN
TUPLE
Vector direction for agent to move
Random-walk moves each agent in a random direction.This
is like brownian in [13],except that speed is also randomized.
(flock DIRECTION)
Parameter
Type
Description
DIRECTION
TUPLE
Preferred direction to flock toward
RETURN
TUPLE
Vector direction for agent to move
The flock behavior,also from [13],moves agent groups
by repelling the closest agents,aligning with moderately-
proximate agents,and weakly attracting distant agents.This
allows a group of agents to “flock” together in partially-
coherent group motion.The DIRECTION argument guides the
motion of the flock with a preferred direction supplied to some
or all members,as investigated in [14].
(flock-to LOCATION)
Parameter
Type
Description
LOCATION
TUPLE
Coordinates to flock to
RETURN
TUPLE
Vector direction for agent to move
(a) Random-Walk
(b) Flock/Flock-To
(c) Disperse/Scatter
(d) Toward
(e) Cluster-By
Fig.1.Examples of agents being controlled by group behavior primitives written in Proto.
The flock-to behavior is like flock,except the agents
move coherently to a location,rather than toward a direction.
(disperse)
Parameter
Type
Description
RETURN
TUPLE
Vector direction for agent to move
The disperse behavior,our last adaptation from [13],
repels agents away from one another with a force proportional
to the inverse square of the distance separating them.
(scatter DIRECTION)
Parameter
Type
Description
DIRECTION
TUPLE
Vector biasing scatter direction
RETURN
TUPLE
Vector direction for agent to move
The scatter behavior is much like disperse,except that
agents do not slow down when they start getting far apart and
they have a directional bias,DIRECTION.
(toward TARGET)
Parameter
Type
Description
TARGET
BOOLEAN
Boolean indicator that is
true if an agent is a target
RETURN
TUPLE
Vector direction for agent to move
The toward behavior finds the direction toward the mean
location of all neighbors with a TARGET property.
(away-from TARGET)
Parameter
Type
Description
TARGET
BOOLEAN
Boolean indicator that is
true if an agent is a target
RETURN
TUPLE
Vector direction for agent to move
The away-from behavior is the inverse of toward.
(cluster-by GROUP-ID)
Parameter
Type
Description
GROUP-ID
INTEGER
Identifier for an agent group
RETURN
TUPLE
Vector direction for agent to move
Finally,cluster-by,sorts agents into groups:all agents
repel each other weakly and are strongly attracted to others
with the same GROUP-ID identifier.This will tend to separate
the group into clusters by identifier,though if the agents are
widely scattered,there may be more than one cluster for any
given identifier.
C.Agent Scripting Library
The last ingredient needed is a means of composing to-
gether these group behavior primitives to form useful agent
behavior scripts,which will typically be much more com-
plicated.Proto’s native approach is one of purely functional
composition—mathematically elegant,but not well suited for
the way that simulation designers often like to think.Instead,
we would like to be able to talk about a script in terms
of concepts like particular groups being assigned particular
behaviors,responding to triggers,or progressing through a
planned sequence one stage at a time.
Technically,Proto’s functional model can already provide
all of these capabilities.The problemis that the code to do so is
often awkward and does not “look” like the kind of state-based
programming that is more familiar for this sort of scripting.
Fortunately,Proto has recently been extended with a capability
for syntactic macros.We use this macro programming facility
to create new syntactic constructs suitable for agent scripting.
The macros transform these new syntactic constructs into
implementation in terms of standard Proto primitives.
Our initial agent scripting library comprises five constructs,
selected as examples of group and individual behavior se-
lection and sequencing;other important categories not yet
included are behavior planning,collective decision making,
etc.In our initial library,group-case and where assign
behaviors to groups of agents,on-trigger sets up a trig-
gered action for a group of agents,priority-list assigns
behavior based on the relative importance of competing pri-
orities,and sequence moves a group of agents through a
planned sequence of actions.All of these are defined with the
assumption that the return value is intended to be a vector field
specifying the movement of agents.We will now detail each
of these constructs in turn,then demonstrate their use with the
examples in the next section.
The syntax of the group-case construct is:
(group-case
(behavior-of MEMBERSHIP-TEST BEHAVIOR
(behavior-of MEMBERSHIP-TEST BEHAVIOR
...
(default BEHAVIOR)...)))
This operates much like an ordinary case statement:each
MEMBERSHIP-TEST must be a boolean-value expression,and
agents use the BEHAVIOR of the first behavior-of case they
match.If an agent is not a member of any group,then it uses
the default group’s BEHAVIOR.
Within each behavior-of construct,there is a special vari-
able in-group defined.Computations for a group’s behavior
normally extend over both agents in the group and agents
outside of the group,allowing agents to react to information
from others outside of their group.The in-group variable is
true only for those agents in the group,and can thus be used
to restrict computation to only within the group.
The where construct is a good way to do such a restriction:
(where TEST BEHAVIOR)
This computes BEHAVIOR over the set of agents where TEST
is true,much like the standard Proto if construct,except that
all other devices default to a tuple of zeros.
The on-trigger construct has identical syntactic structure:
(on-trigger TRIGGER BEHAVIOR)
Its function,however,is to enable a BEHAVIOR in a dormant
group of agents as soon as TRIGGER becomes true for at
least one member of the group.Once enabled,the group stays
enabled and continues to act.
The syntax of the priority-list construct is:
(priority-list
(priority NAME TEST BEHAVIOR
(priority NAME TEST BEHAVIOR
...)))
Each agent walks the list of priorities in descending order,
treating the first entry as highest priority.When a TEST
evalutes to true,the agent executes the associated BEHAVIOR.
If no priority holds,then the agent does nothing.
Finally,the sequence construct,which moves agents
through a sequence of actions over time,is:
(sequence
([stage|group-stage] NAME ACTION TERMINATION
([stage|group-stage] NAME ACTION TERMINATION
...
[end-sequence|repeat])...))
Agents transition individually through stage constructs and
transition collectively out of group-stage constructs.The
sequence begins with the first stage,executing ACTION until
the TERMINATION condition is met.When an agent finishes a
stage,it just moves on and begins executing the next stage’s
ACTION.For a group-stage,on the other hand,the agent
also informs all neighboring agents,which move on to the
next stage and inform their neighbors as well,and so on until
all agents with reach of communication have changed stages.
Thus,a group-stage terminates when any agent in the group
reaches its TERMINATION condition.
When agents reach the final action in the sequence,their
behavior depends on the final keyword.If the keyword is
end-sequence,the agents stop moving;if it is repeat,the
sequence begins again.Optionally,the keyword ongoing may
be substituted for the last stage’s TERMINATION,in which case
the last stage continues indefinitely instead.
Although these five constructs are just a beginning of the
type of constructs that are necessary to make up a full-fledged
agent scripting library,they demonstrate that Proto macros can
allow more “natural” scripting for agent behaviors.
IV.VALIDATION
We now have an agent scripting library written in Proto and
the ability to execute Proto programs in Unity—all of the in-
gredients necessary for validating our approach to lightweight
simulation scripting.In this section,we demonstrate the power
of our approach by constructing three simulations where
groups of agents need to interact and to coordinate their
behaviors with one another.
(a) Red team advances on Blue team
(b) Blue team notices incoming Red team
(c) Blue team scatters
Fig.2.The red-advance script running on 30 agents.
For these simulations,we consider environments with two
teams of agents:“Red team” aggressors and “Blue team”
defenders.In each simulation,agents from both teams are
placed onto a geo-typical terrain where they can then execute
their group behaviors within the physics and terrain based
constraints of the environment.We run these simulations with
10 to 30 agents;since the code is written in Proto,however,the
same simulations can be executed on any number of agents.
A.Red Advances on Blue
We begin with a simple scenario where Red team advances
on Blue team and Blue scatters and flees when Red gets close:
(def red-advance (red-team blue-team)
(group-case
(behavior-of red-team;;Red team behavior:
(where in-group
(flock-to (tup 0 0)));;go to Blue starting location
(behavior-of blue-team;;Blue team behavior:
(on-trigger (can-see red-team);;when Red is near...
(scatter (away-from red-team)));;flee from Red!
(default (tup 0 0))))))
We do this by using the group-case construct to specify
behavior by team.For Red team,we use flock-to to advance
coherently towards the starting location of Blue team.For Blue
team,we use on-trigger to scatter when any Blue agent
notices an advancing Red,using the bias argument to making
sure that the Blue agents move away-from Red.If any agent
is not on either Red or Blue team,it does nothing.Figure 2
shows agents executing this scenario in Unity.
(a) Team moving as a group
(b) Team breaking up into three sub-groups
(c) Sub-groups moving to different destinations
Fig.3.The deploy script running on 10 agents.
B.Red Deploys from a Vehicle
The next scenario has Red teamdeploying out of an armored
transport vehicle into three squads:
(def deploy (squadID)
(sequence
(stage leave-vehicle;;First stage:
(flock (tup -1 0 0));;move left...
(timeout 20);;...for twenty seconds.
(stage group-by-squad;;Second stage:
(cluster-by squadID);;group into squads...
(timeout 50);;...for fifty seconds.
(stage deploy-to-destination;;Third stage:
(group-case;;Each squad goes to a different location:
(behavior-of (= squadID 0);;First squad...
(flock-to (tup 50 100));;...goes to (50,100)
(behavior-of (= squadID 1);;Second squad...
(flock-to (tup -200 0));;...goes to (-200,0)
(behavior-of (= squadID 2);;Third squad...
(flock-to (tup -100 -100));;...goes to (-100,-100)
(default (tup 0 0))))))
ongoing;;Sequence doesn’t end or repeat
end-sequence)))))
Here,we use the sequence construct to break the deployment
into three phases.First,the agents all flock for 20 seconds to
leave the vehicle together.Next,the agents use cluster-by
to sort themselves out into squads,giving 50 seconds for the
squads to organize themselves.Finally,we use group-case to
have each of the three squads flock-to its own destination.
Figure 3 shows agents executing this scenario in Unity.
(a) Blue team on patrol — looking for Red team.
(b) Blue team members break off to chase Red team.
Fig.4.The patrol-encounter script running on 30 agents.
C.Red Tries to Sneak Past a Blue Patrol
Our third scenario is the most complex:Blue team is trying
to defend against Red team while patrolling a regular pattern.
Meanwhile,Red team is trying to pass through the area that
Blue team is guarding without being caught.
We first define the patrol pattern to be used by Blue team:
(def patrol ()
(sequence
(group-stage checkpoint-1;;First stage:
(flock-to (tup 100 50));;Go toward (100,50)...
;;...until somebody in the patrol is within 5 meters of the place...
(< (vlen (- (coord) (tup 100 50))) 5)
(group-stage checkpoint-2;;Second stage:
(flock-to (tup 0 50));;...now go to (0,50)...
(< (vlen (- (coord) (tup 0 50))) 5)
(group-stage checkpoint-3;;Third stage:
(flock-to (tup 50 -50));;...and then to (50,-50)...
(< (vlen (- (coord) (tup 50 -50))) 5)
repeat)))));;...and repeat
Here,we use a repeating sequence construct to define a cyclic
patrol around three checkpoints.For each checkpoint,the
agents use flock-to to move as a group to that checkpoint.
Once any agent in the group reaches the checkpoint,the
group-stage construct means its information will spread,
causing the whole group to head for the next checkpoint even
if some members have not yet reached the current checkpoint.
We then use this script as a behavior in the scripts for the
overall encounter between Red team and Blue team:
(def patrol-encounter (red-team blue-team)
(group-case
(behavior-of red-team;;Red team behaviors
(priority-list
(priority defend-self
(can-see blue-team);;if Blue shows up...
(scatter (away-from blue-team));;...then run away
(priority invade;;otherwise,
(timeout 500);;when the script says start
(flock-to (tup 200 0)))));;try to pass by Blue team
(behavior-of blue-team;;Blue team behaviors
(priority-list
(priority attack-red
(can-see red-team);;if Red shows up...
(let ((dir (toward red-team)));;...then track...
(where in-group (flock dir)));;...and chase them
(default;;otherwise,
(where in-group (patrol)))));;walk your patrol route.
(default (tup 0 0))))))
As before,we use group-case to specify a behavior for each
team.We now also use priority-list to give each team
multiple possible behaviors,depending on circumstances.Red
teambegins to invade the area 500 seconds after the simulation
starts,attempting to flock-to its target.If a Red team agent
encounters Blue team,however,this goal will be pre-empted
by the goal of defending itself,and it will scatter,fleeing
away-from Blue team.
The Blue team has a complementary priority-list
script:when there are no Red team agents nearby,they default
to patrolling on the three-checkpoint pattern that we defined
above.If a Red team agent is nearby,though,they will break
off to attack,using flock to move toward nearby red-team
agents.Figure 4 shows agents executing this scenario in Unity.
Taken together,these demonstrate the power of our ap-
proach to scripting of agent-based simulations.Unity provides
realistic physics simulation,while Proto and our new agent
scripting library allow for compact scripting of scenarios in
which groups of agents engage in many types of interactions.
The scenarios presented are remarkably compact in code,
requiring only 8 lines,19 lines,and 31 lines respectively.The
same scenarios written in a conventional scripting language
would typically take at least an order of magnitude more
code.We can measure this in some cases by comparing
against similar scripts available on the Unity community site,
http://www.unifycommunity.com/.For example,the
Proto flock code presented above takes only 12 lines,while a
Unity JavaScript version takes 77 lines,yet must “cheat” in its
calculations and can only run a single flock.Similarly,a single-
agent waypoint following Program in Unity is implemented
with 65 lines of JavaScript code,while it takes only 7 lines of
Proto code to script coherent waypoint following for groups
of agents.Even single agent functions are often much simpler
in Proto:a Unity C#script for a wandering agent requires 40
lines,while the equivalent Proto function dither (part of the
standard Proto library) requires only 5 lines.
Also of note is the relatively light computational burden
of Proto;in the scenarios presented,the limiting factor on
Unity simulation speed appears to be the cost of rendering the
agents,with the cost of Proto computation and communication
insignificant.
While not yet a definitive study,these results do clearly
indicate that it is reasonable to expect large benefits from
scripting agent-based simulations in Proto.We believe such a
drastic reduction in size is likely to be due to two factors.First,
as a functional programming language,Proto tends to produce
more compact code.More importantly,however,Proto’s ability
to program aggregates and to execute routines on subgroups
makes it just as easy to script a group behavior as a behavior
for a single individual.The relative importance of these two
factors,however,is not yet established.
V.CONTRIBUTIONS
We have presented a novel approach to construction of
agent-based simulation,based on the integration of the Unity
simulation engine with the Proto spatial computing program-
ming language.We have developed a library of agent group
behaviors and scripting constructs aimed at programming this
environment,and have demonstrated that the combination
allows succinct specification of complex simulation scenarios
with large numbers of interacting agents.
While the results presented in this paper demonstrate the
potential for major improvements in agent-based simulation
programming,there is much more that can be accomplished.
Future work includes better integration between Proto and
Unity,refinement and extension of the behavior library and
scripting constructs,and construction of virtual sensors to
allow Proto-controlled agents access to more information
available from the Unity simulator,such as terrain properties,
physical contact and line-of-sight.
REFERENCES
[1] J.Beal and J.Bachrach,“Infrastructure for engineered emergence on
sensor/actuator networks,” IEEE Intelligent Systems,2006.
[2] M.Lewis,J.Jacobson,and C.based Games,“Game engines in scientific
research,” 2002.
[3] T.Susi,M.Johannesson,and P.Backlund,“Serious games – an
overview,” 2007.
[4] T.Hussain,B.Roberts,C.Bowers,J.Cannon-Bowers,E.Menaker,
S.Coleman,C.Murphy,K.Pounds,A.Koenig,R.Wainess,and
J.Lee,“Designing and developing effective training games for the US
Navy,” in 2009 Interservice/Industry Training,Simulation and Education
Conference.,2009.
[5] “Unity — 3D game engine,” Available:http://unity3d.com/,
Retrieved March 4,2012.
[6] J.Beal,S.Dulman,K.Usbeck,M.Viroli,and N.Correll,“Orga-
nizing the aggregate:Languages for spatial computing,” CoRR,vol.
abs/1202.5509,2012.
[7] A.Rao and M.Georgeff,“BDI agents:From theory to practice,” in
Proceedings of the first international conference on multi-agent systems
(ICMAS-95).San Francisco,1995,pp.312–319.
[8] W.C.Regli,I.Mayk,C.J.Dugan,J.B.Kopena,R.N.Lass,P.J.Modi,
W.M.Mongan,J.K.Salvage,and E.A.Sultanik,“Development and
specification of a reference model for agent-based systems,” Trans.Sys.
Man Cyber Part C,vol.39,pp.572–596,September 2009.[Online].
Available:http://portal.acm.org/citation.cfm?id=1656816.1656823
[9] D.N.Nguyen,K.Usbeck,W.M.Mongan,C.T.Cannon,R.N.
Lass,J.Salvage,and W.C.Regli,“A methodology for developing an
agent systems reference architecture,” in 11th International Workshop
on Agent-oriented Software Engineering,Toronto,ON,May 2010.
[10] E.Sklar,“Netlogo,a multi-agent simulation environment,” Artificial life,
vol.13,no.3,pp.303–311,2007.
[11] J.-L.Giavitto,O.Michel,J.Cohen,and A.Spicher,“Computation in
space and space in computation,” Univerite d’Evry,LaMI,Tech.Rep.
103-2004,2004.
[12] K.Usbeck and J.Beal,“An agent framework for agent societies,”
Systems,Programming,Languages and Applications:Software for Hu-
manity,2011.
[13] J.Bachrach,J.Beal,and J.McLurkin,“Composable continuous space
programs for robotic swarms,” Neural Computing and Applications,
vol.19,no.6,pp.825–847,2010.
[14] I.Couzin,J.Krause,N.Franks,and S.Levin,“Effective leadership and
decision-making in animal groups on the move,” Nature,vol.433,no.
7025,pp.513–516,2005.